Backend engineering

At The Parker Avery Group I'm the main author of Ask Udi, an AI assistant that sits on top of retail data. You ask it a question in everyday language and get back a real table or chart, and the answer can easily be thousands of rows deep. The whole design rests on one idea: the language model never handles the bulk of the data. Its only job is to write the database query and run it by calling a tool, and the server executes it. The rest keeps the full result under a specific artifact, and gives the model sample rows, the total row count and a placeholder that stands for the full table. The model writes its answer around that placeholder without ever seeing the whole thing. Then the browser spots the placeholder, asks for the complete result by its id and draws it as a table or a chart. So the model stays cheap and fast, the raw rows never go to it, and you still get every single row. I've been a Senior Developer there since 2022.

WebSonar came from DEGA Partners, a law firm with a brand-protection practice that wanted to take on more work at a bigger scale. Until 2019 their lawyers built brand-violation cases by hand, collecting the offending pages one at a time next to the case, while those pages kept changing under them. The scanner loaded every page in a headless browser and checked its text against the keyword list of the task. A page with no match got dropped. A matching page got saved twice: once as the rendered view a visitor sees, and once as the raw page source, both with the matched words marked. Why both? A site can hide a word from the visitor, but it can't hide it from its own markup. Around every match it also collected the site's contacts: phone numbers, email addresses, links to social networks and the WHOIS registration record. The same headless browser then printed the report of each job as a PDF. And that capture really matters, because a violation is often acted on months later, when the page has changed or is gone, so the lawyer sees the page as it was instead of a filename to trust. A stopped run picked up right where it stopped as well.

At Mobilauto I wrote the bridge that moved taxi orders between Mobilauto's dispatch and partner firms running a third-party dispatch product. Say an order comes in and no local car takes it. The bridge sends it to every eligible partner at the same moment, as a real order on each partner's own system. The first partner to assign a car keeps the ride, and every other copy gets cancelled in the same second. The rider's quoted price holds too, even when another firm drives: the bridge asks the partner for its price on that route, works out the difference from Mobilauto's price and opens the order with that difference already added. And of course there's the edge case nobody plans for. Later another engineer carried most of the bridge, and I came back to it again and again over the years for the order logic, the retries, the watcher that handles failover and the slow parts that needed speed.

On the faculty site voting is the one place where carelessness actually costs something. Voters log in with a Google Workspace account, so every ballot is tied to a real person's account. Each vote is written in one atomic step, which closes the race where two clicks in the same instant could count twice. The voting path is rate limited, and it reads the client IP from behind the trusted proxy, so a voter can't forge it.

Cropman was a browser batch photo editor I built in 2019 with 3 classmates from the DAN.IT course. I led the team and wrote the server almost alone, and all the pixel work ran there. Undo lived on the server as well: a saved project keeps the originals and the ordered list of every operation, and to step back the browser just sends a shorter list for the server to replay, so undo survives a closed browser window. The paid "Cropman Plus" plan went through a monthly LiqPay subscription, and around 100 paying subscribers covered the running cost.